The field of Large Language Models (LLMs) is moving towards improving their performance in multi-turn conversational exchanges. Current research focuses on addressing the challenges of LLMs getting lost in conversations and failing to recover from incorrect assumptions. A significant direction of research is the development of frameworks and methods for evaluating and improving instruction following in LLMs, including the creation of more diverse and realistic benchmarks. Another area of exploration is the use of LLMs as adaptive tutors in language learning, where the goal is to constrain model outputs to be appropriate for a student's competence level. Researchers are also investigating the generalization of LLMs across different conversational formats, including the detection of truthful and false statements. Notable papers in this area include:

• A study that proposes a multi-dimensional constraint framework for evaluating instruction following in LLMs, which achieved substantial gains in performance without degrading general capabilities.
• A paper that explores the generalization of LLM truth directions on conversational formats and proposes a solution to improve generalization to longer conversation formats.